Audit student research and source quality

1

Research

Why this matters, and what good looks like

A source audit connects each research claim to its original evidence and a recorded human check. Corroboration means comparing it with independent evidence. These professional and government guidance excerpts support inquiry and caution about AI accuracy, rather than a complete disciplinary appraisal curriculum. They establish no AI benefit for this workflow. The register and checking sequence below are authorial design choices. The catalogue strength is an editorial classification, not a new evidence rating.

PBLWorks connects open-ended inquiry and attention to quality with finding and using resources.
UNESCO cautions that fluent output cannot be trusted as accurate and retains human agency.
DfE warns that AI can present false information as fact; evidence on learner impacts is limited and settings must assess risks.

Where the evidence comes from

Strong for disciplinary literacy and sustained inquiry; AI benefit limited and risk-sensitive.

2

Workflow

Brief it, steer it, check it

Bound the audit

Bring your inquiry questions and source records, optionally from L04, plus C01 research concerns or equivalent teacher notes. Limit this audit to one claim and two short originals. Remove personal data. Stop and re-scope if work exceeds 60 active minutes.

Prompt 1 · Organise records

CONTEXT
Inquiry and disciplinary appraisal criteria: [inquiry and criteria]
Source records, exact excerpts with locators, and existing register or none: [source records]
Teacher research concerns and scope: [concerns and scope]
REQUEST
Organise only supplied records for one claim and at most two short sources. Identify missing originals, metadata and evidence. Retain the student's exact claim separately from any candidate revision.
QUALITY BAR
Treat records as unverified. Do not browse, invent bibliography or certify credibility. Every AI-suggested claim or citation is UNUSABLE until human checking of originals. Quote excerpts exactly; distinguish copied quotation from candidate paraphrase. If criteria or records are missing, return only HOLD and the missing input. Never include personal data.
FORMAT
Return a table: claim ID | exact student claim | source ID | original locator | author/date as supplied | exact excerpt and locator | missing information | status. Use UNUSABLE for every status, followed by a short teacher-check list.

Prepare human verification

Check the inventory against your records and record corrections. Ask for questions students can take to the originals. You choose the appraisal questions; AI cannot decide whether evidence is trustworthy.

Prompt 2 · Prepare checks

CONTEXT
Exact step 1 output: [step 1 output]
Teacher corrections and approval of inventory for planning only: [inventory decision]
Original records and disciplinary criteria: [records and criteria]
REQUEST
Prepare an original-source checking route for each source. Include authorship, date/context, evidence supporting the claim, relevance, citation accuracy and independent corroboration. Extract one exact quotation and propose one faithful paraphrase, preserving limitations and locators.
QUALITY BAR
Use supplied excerpts only. Do not infer authority from formatting or invent sources to corroborate. If approval or excerpts are missing, return only HOLD and missing inputs. Both the quotation/citation and paraphrase/claim remain UNUSABLE until humans open originals, compare wording/context and corroborate. Ask who produced the evidence and whether the second source is genuinely independent. Do not score credibility.
FORMAT
Return a table: source ID | original-opening route | exact quotation and locator | candidate paraphrase | verification questions | corroboration question | status. Follow with a blank human check form: source ID; original opened; author/date checked; context and evidence checked; quotation/paraphrase fidelity; citation checked; independent comparison and limits; checker record/version; decision.

Record checks and revise

Students open originals and compare evidence; you personally check consequential claims, citations and quotation/paraphrase fidelity. Record independent corroboration and unresolved limits. Supply only nonidentifying check records. A missing, stale or conflicting check leaves that item unusable.

Prompt 3 · Reconcile the register

CONTEXT
Exact step 2 output and original inventory: [step 2 and inventory]
Teacher-selected questions, edits and completed human original-source/corroboration checks, with versions: [human checks]
REQUEST
Update the source-audit register from explicit human records. Preserve the original student claim and show any human-approved revision separately. Connect each quotation or paraphrase to its original locator and citation record.
QUALITY BAR
You organise supplied decisions, never verify truth. Only copy HUMAN-CHECKED where the human record explicitly covers original access, authorship/date, context, evidence, quotation/paraphrase fidelity, citation and independent corroboration with limits. Missing or conflicting checks mean UNUSABLE/HOLD. Keep test-only checks labelled SIMULATED, never real verification. Do not invent replacement citations or resolve contradictions yourself.
FORMAT
Return a table: claim ID | original claim | human-approved revision | source ID and original locator | author/date | relevance | corroboration and limits | quotation/paraphrase and locator | citation record | human check record/version | status. Then list unresolved checks.

Audit the selected register

Choose and edit the register yourself. Compare the final audit with originals, then decide what students may use. AI checks traceability only. Keep unsupported claims on HOLD and retain the human decision record beside your working register.

Prompt 4 · Audit traceability

CONTEXT
Exact step 3 output: [step 3 output]
Teacher-selected register and explicit edits: [selected register]
Original records, criteria and versioned human checks: [audit evidence]
REQUEST
Audit the selected register for omitted checks, changed wording, unsupported status, lost limitations and mismatched citation/quotation/paraphrase locators. Do not rewrite it.
QUALITY BAR
Compare supplied text only. Never certify truth, credibility or permission to use. Every suggested citation or claim without a complete human original-source check remains UNUSABLE. Missing selection or evidence means HOLD. Synthetic verification stays SIMULATED; actual release remains a human decision.
FORMAT
Return a table: item | exact register text | matching supplied record/version or MISSING | discrepancy | human action. End with Unresolved holds and Human release decision required. If no discrepancy is found, state only that no discrepancy was found in the supplied comparison.

Check before you use it

Have humans opened originals and checked every claim, citation, quotation and paraphrase against context?
Does each item retain its author, date, original locator and human check record?
Have students compared independent evidence and recorded disagreement or limits?
Are unchecked, inaccessible, stale or conflicting items still UNUSABLE or HOLD?
Have you removed personal data and retained the teacher's final use decision outside AI?

Stop rule

Every AI-suggested citation or claim is unusable until checked in the original source.

3

Reuse it

So the next one takes minutes

Suitable

Build a skill

Save this as a reusable instruction you can run on any project.

The bounded sequence can be reused with new teacher materials and fresh human decisions.

Not suitable

Build an agent

Chain the steps, with a checkpoint where you approve.

Original evidence checks and consequential support or use decisions require direct human control in this bounded workflow.